[dpdk-dev] [PATCH] rte_mbuf: scattered pktmbufs freeing optimization

[Olivier MATZ]

Hi Vadim,

On 03/07/2015 12:24 AM, Vadim Suraev wrote:
> Hi, Olivier,
> I realized that if local cache for the mempool is enabled and greater
> than 0,
> if, say, the mempool size is X and local cache length is Y (and it is
> not empty,Y>0)
> an attempt to allocate a bulk, whose size is greater than local cache
> size (max) and greater than X-Y (which is the number of entries in the
> ring) will fail.
> The reason is:
> __mempool_get_bulk will check whether the bulk to be allocated is
> greater than mp->cache_size and will fall to ring_dequeue.
> And the ring does not contain enough entries in this case while the sum
> of ring entries + cache length may be greater or equal to the bulk's
> size, so theoretically the bulk could be allocated.
> Is it an expected behaviour? Am I missing something?

I think it's the expected behavior as the code of mempool_get()
tries to minimize the number of tests. In this situation, even if
len(mempool) + len(cache) is greater than the number of requested
objects, we are almost out of buffers, so returning ENOBUF is not
a problem.

If the user wants to ensure that he can allocates at least X buffers,
he can create the pool with:
  mempool_create(X + cache_size * MAX_LCORE)

> By the way, rte_mempool_count returns a ring count + sum of all local
> caches, IMHO it could mislead, even twice.

Right, today rte_mempool_count() cannot really be used for something
else than debug or stats. Adding rte_mempool_common_count() and
rte_mempool_cache_len() may be useful to give the user a better control
(and they will be faster because they won't browse the cache lengths of
all lcores).

But we have to keep in mind that for multi-consumer pools checking the
common_count before retrieving objects is useless because the other
lcores can retrieve objects at the same time.

Regards,
Olivier